Research On Moving Target Detection And Hardware Acceleration Of Small Airborne Platform

Small airborne platform based moving target detection is widely used in military reconnaissance,rescue exploration and other fields.The airborne platform belongs to the typical moving platform motion target detection scene.Both of the imaging platform and the target are in motion,which increases the difficulty of target detection.However,the small airborne platform is also a typical embedded application scenario.Computing resources and storage resources are limited,while it has extremely high requirements for real-time performance and strict constraints on power consumption and volume.Therefore,this paper studies the real-time algorithm of moving targets detection for small airborne platforms,and its hardware acceleration implementation.First of all,this paper selects a moving target detection algorithm based on ORB registration according to the characteristics of the small airborne platform moving target detection task.This method has a good trade-off between detection accuracy and processing speed.However,there is a problem that the accuracy of global motion estimation decrease a lot when the feature points are concentrated around the target.What’s more,when running on a general-purpose PC platform,the processing time of the algorithm still does not meet the high requirements of the small airborne platform.Both of these point to ORB feature extraction and description algorithm.This paper proposes an evaluation method for algorithm processing speed,which is unrelated with the computing platform and design optimization.According to the method,the calculation complexity and storage access conditions of each step of the ORB algorithm are analyzed,and the algorithm bottleneck is located.This paper proposes two strategy to improved the ORB algorithm.The first strategy is called feature selection based region ranking,which makes the distribution of ORB feature points in the whole image more dispersed.This strategy effectively solves the problem of inaccurate global motion estimation,and also greatly reducing the amount of calculation of ranking feature points.This paper also proposes a strategy for offline calculation of the rotation point pair mode,which removes a large number of floating-point multiplication operations and reduces the algorithm processing time.Finally,this paper carries out hardware accelerated design of the improved algorithm model,and proposes a moving target detection architecture based on FPGA+DSP.Then we deploy the architecture on FPGA+DSP heterogeneous computing platform to realize a moving target dector.The experimental results show that the moving target detector maintains a high detection accuracy and the processing time is extremely low.For 640*480 resolution images,the processing frame rate is up to 160 fps.The resource use is little and the power consumption is less than 6W.The proposed detector is suitable for small airborne platforms with limited resources,high real-time requirements and strict power consumption constraints.